Reproducible crash with Boost Asio and io_service::post

[boost-bugs@…]

On OSX, running the following code repeatedly results in an occasional crash with the stack trace following. Unfortunately, I can't find a workaround nor a fix since I don't actually understand why it is crashing...

  using namespace boost::asio;

  io_service service;
  io_service::work work(service);
  boost::thread thread{[&](){
      service.reset();
      service.run();
    }
  };
  service.post([&service](){service.stop();});
  thread.join();

Error: signal 11:
0   KimiTestApp                         0x0000000105534a02 _Z7handleri + 34
1   libsystem_platform.dylib            0x00007fff97a075aa _sigtramp + 26
2   KimiTestApp                         0x00000001054f154d _ZN5boost6detail12shared_countD2Ev + 45
3   libsystem_pthread.dylib             0x00007fff9665691e _pthread_cond_signal + 612
4   KimiTestApp                         0x0000000105642c43 _ZN5boost4asio6detail11posix_event17signal_and_unlockINS1_11scoped_lockINS1_11posix_mutexEEEEEvRT_ + 115
5   KimiTestApp                         0x0000000105642ab4 _ZN5boost4asio6detail15task_io_service31wake_one_idle_thread_and_unlockERNS1_11scoped_lockINS1_11posix_mutexEEE + 100
6   KimiTestApp                         0x000000010564293f _ZN5boost4asio6detail15task_io_service26wake_one_thread_and_unlockERNS1_11scoped_lockINS1_11posix_mutexEEE + 47
7   KimiTestApp                         0x000000010564247e _ZN5boost4asio6detail15task_io_service25post_immediate_completionEPNS1_25task_io_service_operationEb + 206
8   KimiTestApp                         0x000000010564002e _ZN5boost4asio6detail15task_io_service4postIZN42GlobalTestModule_TestNoOpDieImmediate_Test8TestBodyEvE3$_0EEvRT_ + 190
9   KimiTestApp                         0x000000010563ff24 _ZN5boost4asio10io_service4postIZN42GlobalTestModule_TestNoOpDieImmediate_Test8TestBodyEvE3$_0EENS0_12async_resultINS0_12handler_typeIT_FvvEE4typeEE4typeEOS7_ + 68
10  KimiTestApp                         0x000000010563fded _ZN42GlobalTestModule_TestNoOpDieImmediate_Test8TestBodyEv + 125

[Ben Strong <bstrong@…>]

We hit this bug and fixed it with the following patch:

--- a/boost/asio/detail/posix_event.hpp
+++ b/boost/asio/detail/posix_event.hpp
@@ -58,8 +58,8 @@ public:
   {
     BOOST_ASIO_ASSERT(lock.locked());
     signalled_ = true;
-    lock.unlock();
     ::pthread_cond_signal(&cond_); // Ignore EINVAL.
+    lock.unlock();
   }

   // Reset the event.

The issue is that in the task_io_service, the thread_info contains an event, and the thread_info lifetime is protected by the mutex associated with the lock passed into signal_and_unlock, so there is a race where the event is destroyed after the lock.unlock() but before the pthread_cond_signal().

The "Ignore EINVAL" comment suggests that someone hit this issue before. Unfortunately, relying on pthread_cond_signal to return an error (and not crash) is not a sufficient solution.

[boost-bugs@…]

I can confirm that the crash is gone with this change, thanks for posting.

[ikostov@…]

found similar issues ​https://svn.boost.org/trac10/ticket/12690. Patch works for me as well